UC Irvine

Proton exchange membrane (PEM) fuel cell is a promising candidate as a renewable energy source in the future. It is an electrochemical device that directly convert chemical energy in hydrogen fuel to electric energy with water as the only byproduct. In operation, multiple physical processes occur, including electrochemical reaction, heat and mass transfer, liquid water formation and vaporization, which govern fuel cell operation and performance. Various physical models have been developed, which couple the multi-physics with electrochemical reaction kinetics, to predict fuel cell performance for design and control purpose. Machine learning and data-driven approach have received a growing attention in recent years. This study investigates the application of artificial neural network (ANN) machine learning to predict PEM fuel cell performance. A novel four-layered backpropagation ANN is developed to achieve reasonable accurate prediction of fuel cell performance using the large amount (> 1,500) of performance data (I-V curves) obtained from a validated three-dimensional (3D) physical model for various operating conditions, including temperature, anode and cathode relative humidity. Various ANN parameters, including the number of neurons in each hidden layer, the sizes of training data, the activation functions, selection of training data set, are investigated to assess their impacts on prediction accuracy. Simulations show 20 neurons in each hidden layer fit best for our database. The hyperbolic tangent function as the ANN’s activation function in hidden layer performs the best in terms of prediction accuracy because of its high gradient value and smooth gradient profile. For data set selection, randomly selected training data show a good ANN prediction. For selected datasets, the ANN shows capabilities of predicting the I-V curves using incomplete input data information and filtering out noise signals or outliers in the input data set. These results can provide a valuable guidance of using the ANN to help PEM fuel cell experiment in which data for training are limited in most case.